Meta Muse Code: Terminal Coding Agent Enters Beta With Muse Spark 1.2

Meta released Muse Code in beta on August 5, 2026, a coding agent that runs in the terminal ¹. Alongside it, the company published Muse Spark 1.2, the coding-focused model that powers it. Meta frames the release as its next step toward the frontier, saying larger and more capable models are on the way ¹.

Muse Code targets complex software engineering work across large repositories: planning changes, writing code, and validating the results ¹. Meta says it can coordinate multiple persistent subagents for each task ¹. Installation is available for macOS and Linux via curl -fsSL https://dev.meta.ai/install.sh | bash ¹.

Subagents That Stay Alive Across a Session

The most distinctive part of Muse Code’s design is how it treats background agents. By Meta’s account, Muse Code runs a simple agent loop plus a set of async background agents that extend the main agent’s capability. These are not spawned per task — they remain active throughout each session ¹, which Meta says helps avoid gathering the same information repeatedly.

The background agents carry out next steps on their own and decide when to report back to the main agent ¹. Meta says this persistence reduces both latency and the need for steering on difficult, multi-step tasks ¹.

The second design point is the runtime. Muse Code appends every model call, tool run, approval, and edit to a local event log ¹. Meta treats that log as the single source of truth, which in its account leaves the runtime “replay-exact and restart-safe” — a crash costs the agent nothing but the moment, not its place in the work ¹. The target is clearly the long-running task that a mid-run failure would otherwise throw away.

Several skills come bundled, and Meta walks through three. /plan converts a task into a plan that has to clear an approval gate; /grill keeps pressure-testing that plan until it survives; /goal drives toward whatever objective has been set ¹.

Muse Spark 1.2 Is a Coding-Focused Update

Muse Spark 1.2 is a coding-focused update to Muse Spark 1.1, which shipped on the Meta Model API in July. Meta says it improves code generation, complex debugging, codebase understanding, and end-to-end developer workflows ¹. Training compute on coding tasks was scaled up significantly, and the diversity of training environments was expanded ¹.

Of the training details Meta sets out, two are worth pulling out. The first is co-training: Meta brought Muse Spark 1.2 up alongside Muse Code rather than separately. What went into that, per Meta, was harness trajectories filtered by rejection sampling, tuned recipes covering goals, compaction, and subagents, and the Muse Code toolset folded into the mix ¹. The model, in other words, was tuned for the state it ships in — paired with the agent, not standalone.

The second is a self-improvement loop. Meta used Muse Spark 1.1 to generate challenging coding environments and instruction-following templates, then had the same model grade candidate solutions against those requirements to build a training dataset for 1.2 ¹. Meta says this helped 1.2 follow complex instructions more precisely than its predecessor ¹.

As a case study, Meta describes GPU kernel optimization: the model writes, compiles, profiles, and progressively improves kernel performance against a provided baseline over more than 1,000 tool calls, running up to 24 hours ¹. The targets were KDA and MLA kernels for NVIDIA Hopper GPUs, with the FLA Triton implementation of KDA as the baseline. Direct imports of outside kernel libraries — FLA among them — were off limits, so the algorithm had to be built up in Triton rather than wrapped around code that already worked ¹. Meta says the agent continues to achieve substantial improvements over the baseline ¹. The size of that improvement is not given as a number in the body of the announcement.

A Contributor Tier That Trades Training Rights for Price

On cost, the practical detail sits in the Meta Model API price list, which carries two tiers ².

The Standard tier covers muse-spark-1.1 and muse-spark-1.2 at $1.25 per million input tokens, $4.25 per million output tokens, and $0.15 for cached input ². Meta states explicitly that prompts and completions on this tier are not used to train Meta models ².

The Contributor tier covers muse-spark-1.2-contributor at $0.10 per million input tokens, $0.20 per million output tokens, and $0.002 for cached input ². Meta frames this as a trade: steeply cut token rates in return for the right to train later Meta models on what you send and what comes back. The uses it points to are prototyping, checking that an integration works, and scaling up experiments — on the condition that having your data trained on is something you can accept ².

Rate limits differ by tier as well, applied per team: Standard allows 3,000 requests and 4 million tokens per minute, Contributor 100 requests and 3 million tokens per minute ². That gap suggests the Contributor tier is not intended for production workloads. Separately, web search grounding costs $2.50 per 1,000 queries on top of tokens, and there is no long-context premium ².

The practical consequence is that any environment where business code and prompts cannot be used for training is effectively limited to the Standard tier. Comparing AI coding tools requires putting their pricing side by side, but for Muse Code the structure to note first is that the cheaper tier comes with a condition attached.

Billing itself follows the Meta Model API’s usage-based model. Meta’s documentation describes charges based on the tokens your requests consume, not a monthly subscription ³. On first run, the agent asks you to sign in through a browser or paste an API key; scripts and CI use an API key ³, supplied through the META_API_KEY environment variable ³.

Where It Lands Among Terminal-Resident Agents

Meta opened the Meta Model API in public preview with Muse Spark 1.1 on July 9, 2026, making this roughly a one-month gap. Where that release was about offering a model through an API, this one packages the model and the agent together.

The terminal-resident agent format itself is already established, whether through Claude Code’s path from an internal CLI tool or OpenAI’s Codex rollout. 9to5Mac reports that, unlike ChatGPT Codex and Claude Code, Muse Code currently has no dedicated app interface ⁴.

For anyone choosing an agent to run internally, three questions fall out of this release. How far does event-log resumability actually get you when long-running tasks are put into CI? How much difference does persistent background agents make to wait times on a real repository? And can the Contributor tier’s condition be accepted for your own code? For evaluation details, Meta points to a separate report ¹.